写在前面

本文将用一个完整的例子，借助PyTorch代码来实现神经网络的BP反向传播算法，为了承接上篇文章，本文中的例子仍然使用到了SGD随机梯度算法¹。【这是深度学习数学原理专题系列的第二篇文章】

PyTorch的 .data()

首先来看PyTorch官方文档里面对该函数的介绍，翻译过来就是，该方法的功能是以标准的Python数字的形式来返回这个张量的值，这个方法只能用于只包含一个元素的张量，对于其他的张量，请查看方法tolist()，该操作是不可微分的，即不可求导。
简单来讲，就是说PyTorch中的.data()可以将变量(Variable)变为tensor，同时将requires_grad设置为Flase²，代码示例如下：

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
# @Time    : 2021/11/26 12:03
# @Author  : William Baker
# @FileName: demo_torch_data.py
# @Software: PyCharm
# @Blog    : https://blog.csdn.net/weixin_43051346# refer:  https://www.cnblogs.com/Guang-Jun/p/14771157.html
# w.grad.item()是等价于 w.grad.data的
import torch# .data()将变量(Variable)变为tensor,将requires_grad设置为Flase，即不会建立计算图，更新时只修改其数值
a = torch.tensor([1.0], requires_grad=True)
print(a.requires_grad)   # Trueb = a.data
print(b, b.requires_grad)        # 输出为： tensor([1.])  False
print(a.requires_grad)           # Trueprint(a.data.item(), type(a.data.item()))    # 1.0 <class 'float'>
print(a.item(), type(a.item()))              # 1.0 <class 'float'># b2 = a.grad.data
# print(b2, b2.requires_grad)      # AttributeError: 'NoneType' object has no attribute 'data'

PyTorch的 .item()

PyTorch中的.item()方法的功能是，以标准的Python数字的形式来返回这个张量的值。这个方法只能用于只包含一个元素的张量。

该方法返回的结果是普通Python数据类型，自然不能调用backward()方法来进行梯度的反向传播。代码示例如下：

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
# @Time    : 2021/11/26 12:16
# @Author  : William Baker
# @FileName: demo_torch_item.py
# @Software: PyCharm
# @Blog    : https://blog.csdn.net/weixin_43051346# refer:  https://www.cnblogs.com/Guang-Jun/p/14771157.html
import torcha = torch.tensor([1.0], requires_grad=True)c = a.item()
print(c, type(c))      # 输出为：1.0 <class 'float'>

BP with SGD的PyTorch代码实现

我们知道，神经网络的正向传播Forward 是计算损失loss，创建新的计算图；而反向传播Backward 是求偏导进行梯度计算，即损失值loss对权重值w求梯度，计算完梯度后会将其存到变量（比如权重w）里面，存完之后计算图就会得到释放。

Gradient:∂loss∂wGradient: \frac{\partial loss}{\partial w} Gradient:∂w∂loss

铺垫完毕，接下来上代码（还是以线性函数为例），具体的讲解尽在要多详细有多详细的注释中：

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
# @Time    : 2021/11/24 17:37
# @Author  : William Baker
# @FileName: SGD_torch.py
# @Software: PyCharm
# @Blog    : https://blog.csdn.net/weixin_43051346import os
os.environ['KMP_DUPLICATE_LIB_OK'] = 'TRUE'
import torch
import matplotlib.pyplot as pltx_data = [1.0, 2.0, 3.0]
y_data = [2.0, 4.0, 6.0]#tensor中包含data(w)和grad(loss对w求导)
w = torch.tensor([1.0])     # w的初值为1.0
w.requires_grad = True      # 需要计算梯度
print(w.data)
# 学习率 learning rate
lr = 0.01def forward(x):return x * w     # x和w都是tensor类型，数乘def loss(x, y):# 每调用一次loss函数，就把计算图构建出来了y_pred = forward(x)return (y_pred - y) ** 2# print("predict (before training)", 4, forward(4).item())
print('训练前的输入值x：{}, 训练前的预测值：{}\n'.format(4, forward(4).item()))
print("***************************开始训练***************************")    # 训练当达到最优值时，即损失函数梯度grad下降到梯度为0，W的值不再继续迭代epoch_list = []
cost_list = []
for epoch in range(100):for x, y in zip(x_data, y_data):# 下面两行，构建计算图的时候，直接使用张量进行运算（但是在权重更新的时候，要使用.data)l = loss(x, y)      # l是一个张量tensor，tensor主要是在建立计算图 forward, compute the loss，Forward 前馈是计算损失loss，创建新的计算图l.backward()        # backward,compute grad for Tensor whose requires_grad set to True 反向传播过程就会自动计算所需要的梯度# Backward 反馈是计算梯度，计算完梯度后会将其存到变量（比如权重w）里面，存完之后计算图就会得到释放# 每进行一次反向传播，把图释放，准备进行下一次的图"""# print('\tgrad:', x, y, w.grad.item(), w.grad.item().type)    # 报错 AttributeError: 'float' object has no attribute 'type'print('\tgrad:', x, y, w.grad.item(), type(w.grad.item()))     # w.grad.item()   # grad:  2.0 4.0 -7.840000152587891 <class 'float'># print('\tgrad-:', x, y, w.grad, type(w.grad))                # w.grad          # grad-: 2.0 4.0 tensor([-7.8400]) <class 'torch.Tensor'>print('\tgrad-*:', x, y, w.grad, w.grad.type())                # w.grad          # grad-*: 2.0 4.0 tensor([-7.8400]) torch.FloatTensorprint('\tgrad--:', x, y, w.grad.data, w.grad.data.type())      # w.grad.data     # grad--: 2.0 4.0 tensor([-7.8400]) torch.FloatTensor"""print('\tgrad:', x, y, w.grad.item())                           # grad: 2.0 4.0 -7.840000152587891# print('\tgrad--:', x, y, w.data.item(), type(w.data.item()))    # grad--: 2.0 4.0 1.0199999809265137 <class 'float'># w -= lr * grad_val        # w = w - lr * gradient(w)   梯度下降的核心所在# print(w.data.requires_grad)    # Falsew.data = w.data - lr * w.grad.data       # 权重更新时，需要用到标量，注意grad也是一个tensor   # w.grad.item()是等价于 w.grad.data的，都是不建立计算图# print(w.data.requires_grad)    # Falsew.grad.data.zero_()     # after update, remember set the grad to zero     # 把权重里面的梯度数据清0，不然就变成了梯度累加epoch_list.append(epoch)cost_list.append(l)# print('progress:', epoch, l.item())print('Progress: Epoch {}, loss:{}'.format(epoch, l.item()))    # 取出loss使用l.item，不要直接使用l（l是tensor会构建计算图）# Progress: Epoch 99, loss:9.094947017729282e-13# print('Progress-: Epoch {}, loss:{}'.format(epoch, l.data.item()))  # Progress-: Epoch 99, loss:9.094947017729282e-13print("***************************训练结束***************************\n")
# print("predict (after training)", 4, forward(4).item())
print('训练后的输入值x：{}, 训练后的预测值：{}'.format(4, forward(4).item()))# 绘图
plt.plot(epoch_list, cost_list)
plt.ylabel('Loss')
plt.xlabel('Epoch')
plt.show()

输出结果如下：


tensor([1.])
训练前的输入值x：4, 训练前的预测值：4.0***************************开始训练***************************grad: 1.0 2.0 -2.0grad: 2.0 4.0 -7.840000152587891grad: 3.0 6.0 -16.228801727294922
Progress: Epoch 0, loss:7.315943717956543grad: 1.0 2.0 -1.478623867034912grad: 2.0 4.0 -5.796205520629883grad: 3.0 6.0 -11.998146057128906
Progress: Epoch 1, loss:3.9987640380859375grad: 1.0 2.0 -1.0931644439697266grad: 2.0 4.0 -4.285204887390137grad: 3.0 6.0 -8.870372772216797
Progress: Epoch 2, loss:2.1856532096862793grad: 1.0 2.0 -0.8081896305084229grad: 2.0 4.0 -3.1681032180786133grad: 3.0 6.0 -6.557973861694336
Progress: Epoch 3, loss:1.1946394443511963grad: 1.0 2.0 -0.5975041389465332grad: 2.0 4.0 -2.3422164916992188grad: 3.0 6.0 -4.848389625549316
Progress: Epoch 4, loss:0.6529689431190491grad: 1.0 2.0 -0.4417421817779541grad: 2.0 4.0 -1.7316293716430664grad: 3.0 6.0 -3.58447265625
Progress: Epoch 5, loss:0.35690122842788696grad: 1.0 2.0 -0.3265852928161621grad: 2.0 4.0 -1.2802143096923828grad: 3.0 6.0 -2.650045394897461
Progress: Epoch 6, loss:0.195076122879982grad: 1.0 2.0 -0.24144840240478516grad: 2.0 4.0 -0.9464778900146484grad: 3.0 6.0 -1.9592113494873047
Progress: Epoch 7, loss:0.10662525147199631grad: 1.0 2.0 -0.17850565910339355grad: 2.0 4.0 -0.699742317199707grad: 3.0 6.0 -1.4484672546386719
Progress: Epoch 8, loss:0.0582793727517128grad: 1.0 2.0 -0.1319713592529297grad: 2.0 4.0 -0.5173273086547852grad: 3.0 6.0 -1.070866584777832
Progress: Epoch 9, loss:0.03185431286692619grad: 1.0 2.0 -0.09756779670715332grad: 2.0 4.0 -0.3824653625488281grad: 3.0 6.0 -0.7917022705078125
Progress: Epoch 10, loss:0.017410902306437492grad: 1.0 2.0 -0.07213282585144043grad: 2.0 4.0 -0.2827606201171875grad: 3.0 6.0 -0.5853137969970703
Progress: Epoch 11, loss:0.009516451507806778grad: 1.0 2.0 -0.053328514099121094grad: 2.0 4.0 -0.2090473175048828grad: 3.0 6.0 -0.43272972106933594
Progress: Epoch 12, loss:0.005201528314501047grad: 1.0 2.0 -0.039426326751708984grad: 2.0 4.0 -0.15455150604248047grad: 3.0 6.0 -0.3199195861816406
Progress: Epoch 13, loss:0.0028430151287466288grad: 1.0 2.0 -0.029148340225219727grad: 2.0 4.0 -0.11426162719726562grad: 3.0 6.0 -0.23652076721191406
Progress: Epoch 14, loss:0.0015539465239271522grad: 1.0 2.0 -0.021549701690673828grad: 2.0 4.0 -0.08447456359863281grad: 3.0 6.0 -0.17486286163330078
Progress: Epoch 15, loss:0.0008493617060594261grad: 1.0 2.0 -0.01593184471130371grad: 2.0 4.0 -0.062453269958496094grad: 3.0 6.0 -0.12927818298339844
Progress: Epoch 16, loss:0.00046424579340964556grad: 1.0 2.0 -0.011778593063354492grad: 2.0 4.0 -0.046172142028808594grad: 3.0 6.0 -0.09557533264160156
Progress: Epoch 17, loss:0.0002537401160225272grad: 1.0 2.0 -0.00870823860168457grad: 2.0 4.0 -0.03413581848144531grad: 3.0 6.0 -0.07066154479980469
Progress: Epoch 18, loss:0.00013869594840798527grad: 1.0 2.0 -0.006437778472900391grad: 2.0 4.0 -0.025236129760742188grad: 3.0 6.0 -0.052239418029785156
Progress: Epoch 19, loss:7.580435340059921e-05grad: 1.0 2.0 -0.004759550094604492grad: 2.0 4.0 -0.018657684326171875grad: 3.0 6.0 -0.038620948791503906
Progress: Epoch 20, loss:4.143271507928148e-05grad: 1.0 2.0 -0.003518819808959961grad: 2.0 4.0 -0.0137939453125grad: 3.0 6.0 -0.028553009033203125
Progress: Epoch 21, loss:2.264650902361609e-05grad: 1.0 2.0 -0.00260162353515625grad: 2.0 4.0 -0.010198593139648438grad: 3.0 6.0 -0.021108627319335938
Progress: Epoch 22, loss:1.2377059647405986e-05grad: 1.0 2.0 -0.0019233226776123047grad: 2.0 4.0 -0.0075397491455078125grad: 3.0 6.0 -0.0156097412109375
Progress: Epoch 23, loss:6.768445018678904e-06grad: 1.0 2.0 -0.0014221668243408203grad: 2.0 4.0 -0.0055751800537109375grad: 3.0 6.0 -0.011541366577148438
Progress: Epoch 24, loss:3.7000872907810844e-06grad: 1.0 2.0 -0.0010514259338378906grad: 2.0 4.0 -0.0041217803955078125grad: 3.0 6.0 -0.008531570434570312
Progress: Epoch 25, loss:2.021880391112063e-06grad: 1.0 2.0 -0.0007772445678710938grad: 2.0 4.0 -0.0030469894409179688grad: 3.0 6.0 -0.006305694580078125
Progress: Epoch 26, loss:1.1044940038118511e-06grad: 1.0 2.0 -0.0005745887756347656grad: 2.0 4.0 -0.0022525787353515625grad: 3.0 6.0 -0.0046634674072265625
Progress: Epoch 27, loss:6.041091182851233e-07grad: 1.0 2.0 -0.0004248619079589844grad: 2.0 4.0 -0.0016651153564453125grad: 3.0 6.0 -0.003444671630859375
Progress: Epoch 28, loss:3.296045179013163e-07grad: 1.0 2.0 -0.0003139972686767578grad: 2.0 4.0 -0.0012311935424804688grad: 3.0 6.0 -0.0025491714477539062
Progress: Epoch 29, loss:1.805076408345485e-07grad: 1.0 2.0 -0.00023221969604492188grad: 2.0 4.0 -0.0009107589721679688grad: 3.0 6.0 -0.0018854141235351562
Progress: Epoch 30, loss:9.874406714516226e-08grad: 1.0 2.0 -0.00017189979553222656grad: 2.0 4.0 -0.0006742477416992188grad: 3.0 6.0 -0.00139617919921875
Progress: Epoch 31, loss:5.4147676564753056e-08grad: 1.0 2.0 -0.0001270771026611328grad: 2.0 4.0 -0.0004978179931640625grad: 3.0 6.0 -0.00102996826171875
Progress: Epoch 32, loss:2.9467628337442875e-08grad: 1.0 2.0 -9.393692016601562e-05grad: 2.0 4.0 -0.0003681182861328125grad: 3.0 6.0 -0.0007610321044921875
Progress: Epoch 33, loss:1.6088051779661328e-08grad: 1.0 2.0 -6.937980651855469e-05grad: 2.0 4.0 -0.00027179718017578125grad: 3.0 6.0 -0.000560760498046875
Progress: Epoch 34, loss:8.734787115827203e-09grad: 1.0 2.0 -5.125999450683594e-05grad: 2.0 4.0 -0.00020122528076171875grad: 3.0 6.0 -0.0004177093505859375
Progress: Epoch 35, loss:4.8466972657479346e-09grad: 1.0 2.0 -3.790855407714844e-05grad: 2.0 4.0 -0.000148773193359375grad: 3.0 6.0 -0.000308990478515625
Progress: Epoch 36, loss:2.6520865503698587e-09grad: 1.0 2.0 -2.8133392333984375e-05grad: 2.0 4.0 -0.000110626220703125grad: 3.0 6.0 -0.0002288818359375
Progress: Epoch 37, loss:1.4551915228366852e-09grad: 1.0 2.0 -2.09808349609375e-05grad: 2.0 4.0 -8.20159912109375e-05grad: 3.0 6.0 -0.00016880035400390625
Progress: Epoch 38, loss:7.914877642178908e-10grad: 1.0 2.0 -1.5497207641601562e-05grad: 2.0 4.0 -6.103515625e-05grad: 3.0 6.0 -0.000125885009765625
Progress: Epoch 39, loss:4.4019543565809727e-10grad: 1.0 2.0 -1.1444091796875e-05grad: 2.0 4.0 -4.482269287109375e-05grad: 3.0 6.0 -9.1552734375e-05
Progress: Epoch 40, loss:2.3283064365386963e-10grad: 1.0 2.0 -8.344650268554688e-06grad: 2.0 4.0 -3.24249267578125e-05grad: 3.0 6.0 -6.580352783203125e-05
Progress: Epoch 41, loss:1.2028067430946976e-10grad: 1.0 2.0 -5.9604644775390625e-06grad: 2.0 4.0 -2.288818359375e-05grad: 3.0 6.0 -4.57763671875e-05
Progress: Epoch 42, loss:5.820766091346741e-11grad: 1.0 2.0 -4.291534423828125e-06grad: 2.0 4.0 -1.71661376953125e-05grad: 3.0 6.0 -3.719329833984375e-05
Progress: Epoch 43, loss:3.842615114990622e-11grad: 1.0 2.0 -3.337860107421875e-06grad: 2.0 4.0 -1.33514404296875e-05grad: 3.0 6.0 -2.86102294921875e-05
Progress: Epoch 44, loss:2.2737367544323206e-11grad: 1.0 2.0 -2.6226043701171875e-06grad: 2.0 4.0 -1.049041748046875e-05grad: 3.0 6.0 -2.288818359375e-05
Progress: Epoch 45, loss:1.4551915228366852e-11grad: 1.0 2.0 -1.9073486328125e-06grad: 2.0 4.0 -7.62939453125e-06grad: 3.0 6.0 -1.430511474609375e-05
Progress: Epoch 46, loss:5.6843418860808015e-12grad: 1.0 2.0 -1.430511474609375e-06grad: 2.0 4.0 -5.7220458984375e-06grad: 3.0 6.0 -1.1444091796875e-05
Progress: Epoch 47, loss:3.637978807091713e-12grad: 1.0 2.0 -1.1920928955078125e-06grad: 2.0 4.0 -4.76837158203125e-06grad: 3.0 6.0 -1.1444091796875e-05
Progress: Epoch 48, loss:3.637978807091713e-12grad: 1.0 2.0 -9.5367431640625e-07grad: 2.0 4.0 -3.814697265625e-06grad: 3.0 6.0 -8.58306884765625e-06
Progress: Epoch 49, loss:2.0463630789890885e-12grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 50, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 51, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 52, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 53, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 54, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 55, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 56, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 57, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 58, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 59, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 60, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 61, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 62, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 63, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 64, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 65, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 66, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 67, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 68, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 69, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 70, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 71, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 72, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 73, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 74, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 75, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 76, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 77, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 78, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 79, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 80, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 81, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 82, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 83, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 84, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 85, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 86, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 87, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 88, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 89, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 90, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 91, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 92, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 93, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 94, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 95, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 96, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 97, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 98, loss:9.094947017729282e-13grad: 1.0 2.0 -7.152557373046875e-07grad: 2.0 4.0 -2.86102294921875e-06grad: 3.0 6.0 -5.7220458984375e-06
Progress: Epoch 99, loss:9.094947017729282e-13
***************************训练结束***************************训练后的输入值x：4, 训练后的预测值：7.999998569488525Process finished with exit code 0

迭代过程中，Loss值随着Epoch的变化如上图所示，从该图以及程序的输出结果以及可以得出，随着训练逐渐达到最优值，损失函数减低的同时，损失函数对权重的梯度grad也逐渐下降为0，这正是梯度下降到了最低点，即函数梯度值为0的点。如下图二维和三维的图所示，最后来到了红色的点，该点即为要求的最优点³（下面两个图片来自前辈的文章，因为画得已经很好了，懒得再画了，已做引用处理，侵删）。

写到这里，差不多本文也就要结束了。如果我的这篇文章帮助到了你，那我也会感到很高兴，一个人能走多远，在于与谁同行。

参考文章

《PyTorch深度学习实践》完结合集 - 04.反向传播
↩︎
pytorch中.data()与.item()
↩︎
梯度下降算法原理以及其实现
↩︎

使用随机梯度下降SGD的BP反向传播算法的PyTorch代码实现相关推荐

批量梯度下降(BGD)、随机梯度下降(SGD)以及小批量梯度下降(MBGD)的理解
批量梯度下降(BGD).随机梯度下降(SGD)以及小批量梯度下降(MBGD)的理解 </h1><div class="clear"></div> ...
随机梯度下降(SGD)与经典的梯度下降法的区别
随机梯度下降(SGD)与经典的梯度下降法的区别经典的优化方法,例如梯度下降法,在每次迭代过程中需要使用所有的训练数据,这就给求解大规模数据优化问题带来挑战. 知识点:随机梯度下降法(SGD).小批量 ...
批量梯度下降(BGD)、随机梯度下降(SGD)以及小批量梯度下降(MBGD)及 batch、epoch、iteration
先介绍一下梯度下降:梯度下降是一种用于机器学习训练参数的一种优化方法.对损失函数进行梯度下降,"梯度"指误差梯度或误差斜率,"下降"指沿着误差斜率移动到误差较小 ...
BP反向传播算法的思考和直观理解 -卷积小白的随机世界
https://www.toutiao.com/a6690831921246634504/ 2019-05-14 18:47:24 本篇文章,本来计划再进一步完善对CNN卷积神经网络的理解,但在对卷积 ...
梯度的直观理解_BP反向传播算法的思考和直观理解 -卷积小白的随机世界
本篇文章,本来计划再进一步完善对CNN卷积神经网络的理解,但在对卷积层反向传播算法的理解中,越发觉得之前对于BP反向传播算法的理解是不到位的.小白近日觉得,对于深度神经网络,"反向传播&qu ...
通俗理解神经网络BP反向传播算法
转载自通俗理解神经网络BP反向传播算法通俗理解神经网络BP反向传播算法在学习深度学习相关知识,无疑都是从神经网络开始入手,在神经网络对参数的学习算法bp算法,接触了很多次,每一次查找资料学习 ...
[AI 笔记] BP反向传播算法
[AI 笔记] BP反向传播算法说明 BP算法链式法则计算图简单例子前向传播反向传播聚焦计算节点,深入理解计算节点的代码实现例子各种门的功能多元链式法则向量化BP算法参考资料 ...
深度学习中的随机梯度下降(SGD)简介
随机梯度下降(Stochastic Gradient Descent, SGD)是梯度下降算法的一个扩展. 机器学习中反复出现的一个问题是好的泛化需要大的训练集,但大的训练集的计算代价也更大.机器学习 ...
python实现随机梯度下降(SGD)
使用神经网络进行样本训练,要实现随机梯度下降算法.这里我根据麦子学院彭亮老师的讲解,总结如下,(神经网络的结构在另一篇博客中已经定义): def SGD(self, training_data, ep ...

使用随机梯度下降SGD的BP反向传播算法的PyTorch代码实现

Index 目录索引